| title: TemporalMesh Transformer Demo | |
| emoji: 🕸️ | |
| colorFrom: blue | |
| colorTo: purple | |
| sdk: gradio | |
| sdk_version: "5.25.0" | |
| app_file: app.py | |
| pinned: true | |
| license: mit | |
| short_description: Dynamic graph + temporal decay + adaptive depth | |
| tags: | |
| - transformer | |
| - graph-neural-network | |
| - attention | |
| - nlp | |
| - efficient-inference | |
| - adaptive-depth | |
| - mesh-attention | |
| - temporal-decay | |
| - language-model | |
| - pytorch | |
| - visualization | |
| - demo | |
| # TemporalMesh Transformer — Interactive Demo | |
| Visualise **dynamic graph attention**, **temporal decay**, and **per-token adaptive depth routing** on any sentence. | |
| ## What This Demo Shows | |
| - **Exit Gate Heatmap** — which tokens freeze early vs. go deep | |
| - **Dynamic Attention Graph** — how the kNN mesh evolves across layers | |
| - **Token Compute Depth** — actual layers used per word | |
| ## Links | |
| - 📄 [Paper (Zenodo)](https://doi.org/10.5281/zenodo.20287390) | |
| - 🤗 [Model Card](https://huggingface.co/vigneshwar234/TemporalMesh-Transformer) | |
| - 💻 [GitHub](https://github.com/vignesh2027/TemporalMesh-Transformer) | |
| - 📊 [Benchmark Dataset](https://huggingface.co/datasets/vigneshwar234/TMT-Benchmarks) | |